NeST: Network Storage Technologies

1
NeST Network Storage Technologies

• Building an I/O Appliance on Commodity Systems
• John Bent, Remzi Arpaci-Dusseau and Miron Livny

www.nestproject.com
2
Problem Statement
Appliances are attractive because they are
robust, reliable, available and especially
because they are easy to use. To fulfill these
criteria, traditional network appliances impose
policy decisions on their users and are built
either as kernel modules or upon specially
designed kernels.
How to build a portable, configurable I/O
appliance?
3
Goals

• Provide stable, robust and easy to use network
  storage on a variety of platforms
• Allow configuration of server modules to fit
  application needs
• Maintain high server throughput under heavy load
• Degrade client performance gracefully as
  concurrency increases

4
Nest Structure
Protocols
Name Space
Static Configuration
Administrative Interface
Concurrency Architectures
Runtime Adaptation
Storage Management
Consistency Semantics
5
Protocols

• Nest currently supports the following client
  requests in nest-speak
• Send/receive file
• Send/receive partial file
• Query statistics
• Directory operations
• Remove file
• Timestamp request
• Future support ftp, http, nfs, WiND fs.

6
Concurrency Architectures

• Non-blocking single process (NOB)
• Pre-allocated pool of processes (POP)
• Pre-allocated pool of threads (POT)

All concurrency architectures use one control
thread to accept and dispatch client requests.
This approach will more easily allow centralized
decision making as regards admission control,
authentication and quality of services guarantees.
7
Other Modules

• Static configuration
• Build the most appropriate I/O appliance based on
  knowledge of the host system and the application
  needs
• Namespace
• Current flat namespace
• Future hierarchical
• Runtime adaptation
• Bandwidth monitoring and other flows of
  information may prove useful in consistently
  maximizing system throughput

8
Even More Modules

• Consistency semantics
• Well defined behavior is necessary in order to
  allow concurrent client access to the same file
• Storage management
• Current ext2 FS on linux
• Future raw disk management
• Administrative interface
• An easy to use tool must have a nice
  administrative interface. This will probably be
  added through the http protocol.

9
Performance
Server bandwidth in each of the concurrency
architectures remains constant as the number of
clients increases. Client latency and bandwidth
are also affected correctly as client concurrency
grows.
10
Future

• Future Work
• Quality of service
• Admission control, authentication
• Hybrid concurrency architectures
• Static configuration, runtime adaptation,
  namespacing, storage management, consistency
  semantics

• Future Applications
• Condor checkpoint server
• Local staging area for Condor jobs which have
  flocked over the WAN
• Storage brick for WiND

Lots and lots and lots and lots more